1 / 25

Why does subsequence time-series clustering produce sine waves?

Why does subsequence time-series clustering produce sine waves?. IBM Tokyo Research Lab. Tsuyoshi Idé. Contents. What is subsequence time-series clustering? Describing the dependence between subsequences Reducing k-means to eigen problem Deriving sine waves Experiments

sigourney-sandoval
Télécharger la présentation

Why does subsequence time-series clustering produce sine waves?

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript

  1. Why does subsequence time-series clustering produce sine waves? IBM Tokyo Research Lab. Tsuyoshi Idé | 2006/09/19| PKDD 2006

  2. Contents • What is subsequence time-series clustering? • Describing the dependence between subsequences • Reducing k-means to eigen problem • Deriving sine waves • Experiments • beat waves in k-means ! • Summary | 2006/09/19 | PKDD 2006

  3. What is subsequence time-series clustering (STSC)? | 2006/09/19 | PKDD 2006

  4. What’s STSC： k-means clustering of subsequences generated from a time series. Cluster centers are patterns discovered. • subsequences generated by sliding window techniques • subsequences are treated as independent data objects in k-means clustering Cluster centers (centroids) (the average of cluster members) extracted pattern | 2006/09/19 | PKDD 2006

  5. Shocking report Keogh-Lin-Truppel, “Clustering of time series subsequences is meaningless”, ICDM ’03 k-means STSC almost always produces sinusoid cluster centers almost independent of the input time series almost no relation to the original patterns What’s sinusoid effect： unexpectedly, cluster centers in STSC become sinusoids. The reason is unknown. Example concatenate to produce a long time series Explaining why is an open problem k-means STSC Sinusoid cluster centers ! We focus on explaining why. | 2006/09/19 | PKDD 2006

  6. Describing the dependence between subsequences | 2006/09/19 | PKDD 2006

  7. In reality, the subsequences are NOT independent at all. We need to describe the dependence. • subsequences generated by sliding window techniques • subsequences are treated asindependent data objects in clustering ? Let us study how the subsequences are dependent. | 2006/09/19 | PKDD 2006

  8. Theoretical model for time series: Think of a time series as a “state” on a periodic ring. • Assign the valuexlon each lattice points (sites)l. • Attach the orthonormal basis to the sites. • Think of the time series as an n-D vector. Whole time series Artificially assume the periodic boundary condition (PBC) | 2006/09/19 | PKDD 2006

  9. Each subsequence sp is concisely expressed using the translation operator n 1 w-dim vector w: window size “Make p steps backward and take the sites from 1st thru w-th” Definition of the translation operator Ex. shifts e2 with l steps | 2006/09/19 | PKDD 2006

  10. Reducing k-means to eigen problem Easer to analyze theoretically Simple but theoretically a little difficult to handle | 2006/09/19 | PKDD 2006

  11. centroid Rewriting the objective of k-means using the indicator and the density matrix. The objective function of k-means clustering objective to find m (j) Inserting the def of the centroid, and introducing an indicator u(j) as We finally get objective to find u (j). “density matrix” | 2006/09/19 | PKDD 2006

  12. H = … So what? Minimizing E is equivalent to eigen equation. Our goal is to solve it and to show the solution to be sinusoidal. Minimizing E is equivalent to the eigen equation: From the definition, it can be shown eigenvector centroid Let us study the sinusoid effect as ’s eigen equation. | 2006/09/19 | PKDD 2006

  13. Deriving sine waves By solving the eigen equation | 2006/09/19 | PKDD 2006

  14. Mathematical feature of . The expression based on implies a translational symmetry. Fourier basis will simplify the problem. By using , the rho matrix can be written as Summation of shifted ones This form suggests a (pseudo-) translational symmetry of the problem. Translational invariant basis would be more natural So, use the Fourier representation instead of the site representation Window size wave number orthogonal transform | 2006/09/19 | PKDD 2006

  15. When represented in the Fourier basis, is almost diagonal. Thus, the eigenstate is almost pure sinusoid. If we take { fq } as the basis, it follows (after straightforward calculations) power of a Fourier component fq Will be small when a fq is dominant. Theorem When a |fq| is dominant, the eigen state is well approximated by the sine waves with the wavelength of w/|q|, irrespective of the details of the input time series data. | 2006/09/19 | PKDD 2006

  16. Experiments | 2006/09/19 | PKDD 2006

  17. Let us see the correspondence between the two formulations using standard data sets. • eigenvectors minimize the SoS objective • eigenvector  centroid • dominant F.c. governs the eigenvectors spectral STSC k-means STSC k-means STSC | 2006/09/19 | PKDD 2006

  18. [1/2] For data with no particular periodicities, the power concentrates at the longest wavelength w. Only this peak does matter in the centroids. DFT K-means STSC centroids (k=3, w=128) Spectral STSC centroids The resulting centroids are almost independent of the tail of the spectrum. | 2006/09/19 | PKDD 2006

  19. [2/2] STSC centroids become beat waves when a few neighboring |q| are dominant (k=2, w=60). Concatenate 100 instances for each to make a long time series K-means STSC centroids (k=2, w=60) Spectral STSC centroids |q| = 4, 5, 6 are dominant Resulting sine waves exhibit beat wave by interference | 2006/09/19 | PKDD 2006

  20. Summary | 2006/09/19 | PKDD 2006

  21. Summary • The sinusoid effect is an important open problem in data mining. • The pseudo-translational symmetry introduced by the sliding window technique is the origin of the sinusoid effect. • In particular, if there is no particular periodicities within the window size, the clustering centers will be the sine waves of wavelength of w, irrespective of the details of the data. Thank you. | 2006/09/19 | PKDD 2006

  22. Appendix | 2006/09/19 | PKDD 2006

  23. STSC can produce useful results IF some of the conditions of the sinusoid effect are NOT satisfied. For example, if STSC is done locally… A STSC-based change-point detection method (singular spectrum transformation [Ide, SDM’05]) time Cluster center at present Cluster centers in the past comparison to get the CP score | 2006/09/19 | PKDD 2006

  24. In this case, the locality of STSC leads to the loss of (pseudo) translational symmetry , resulting non-meaningless results. SST can produce useful CP detection results, which do depend on the input signal. One general rule could be… “Break the pseudo-translational symmetry” | 2006/09/19 | PKDD 2006

  25. Even for the random data, centroids become sinusoid when w = n (k=2, w=60 and 6000). k=2, w=n=6000(subseq = total) Implication The k-means STSC introduces a mathematical artifact. It is so strong that the resulting centroids are dominated by it. | 2006/09/19 | PKDD 2006

More Related